CN119762161A

CN119762161A - ASA advertisement keyword and material generation system based on AI

Info

Publication number: CN119762161A
Application number: CN202411821638.1A
Authority: CN
Inventors: 罗峰
Original assignee: Beijing Sunshine Weiye Network Technology Co ltd
Current assignee: Beijing Sunshine Weiye Network Technology Co ltd
Priority date: 2024-12-11
Filing date: 2024-12-11
Publication date: 2025-04-04
Anticipated expiration: 2044-12-11
Also published as: CN119762161B

Abstract

This paper proposes an AI-based ASA advertising keyword and material generation system, which uses knowledge graph combined with XLNet to achieve keyword generation, and designs a multimodal Transformer encoder to process advertising materials. The system uses multi-task learning effect estimation and layered multi-armed bandit-based optimization delivery to achieve end-to-end automated processing from advertising demand understanding to delivery optimization.

Description

ASA advertisement keyword and material generation system based on AI

Technical Field

The invention relates to the technical field of advertisements, in particular to an ASA advertisement keyword and material generation system based on AI.

Background

ASA (Apple Search Ads) advertisements are one of the important channels for mobile application promotion. Compared with other mobile advertisement forms, ASA advertisements are directly put on the search result page of Apple App Store, so that the ASA advertisements have high scene correlation and pertinence, and are sharp instruments for reaching high-intention users and improving popularization effects. However, the ASA advertisement has wide keyword coverage and great difficulty in rendering, and how an advertiser generates keywords and advertisement creatives which are highly matched with products rapidly and accurately based on massive material libraries and product information, so that the advertisement effect is improved, and the ASA advertisement optimizing method is a key technical challenge.

Traditional ASA advertisement optimization relies mainly on manually configuring keywords and designing materials. Advertisement optimizers need to have deep understanding of products and industries and sum up targeted keyword combinations and creative strategies to generate high-quality advertisement content. The process is time-consuming and labor-consuming, is easily influenced by experience limitation and cognitive blind areas, and the generated advertisements are often insufficient in generalization capability and difficult to flexibly cope with changeable market environments and user demands. In addition, the screening of keywords and the design of creative lack of quantization indexes and real-time feedback, and advertisers cannot accurately estimate the throwing effect of newly added keywords and materials, so that the utilization rate of advertisement budget is low, and the ROI is difficult to guarantee.

In recent years, the development of artificial intelligence technology brings new break-through for ASA advertisement optimization. Large-scale language models (LargeLanguageModels, LLMs) such as BERT, GPT and the like in the NLP field have made remarkable progress in understanding semantics and generating text. The multi-mode machine learning model is also mature gradually, and can process different forms of material information such as texts, images, videos and the like, so that the computer can 'see' the advertisement content. Meanwhile, the application of the artificial intelligence technology such as knowledge graph, reinforcement learning and the like in the actual scene is also becoming wider. The technological progress of these fronts provides a new idea and method for ASA advertisement generation and optimization.

According to the method, three core problems in ASA advertisement keyword and material generation are developed, namely 1) how to automatically generate ASA advertisement keywords which are highly matched with products based on product information and historical advertisement delivery data so as to improve advertisement effects, 2) how to accurately predict the delivery effects (such as click rate, conversion rate and the like) of new keywords by combining the historical expression data of the keywords, 3) how to automatically generate advertisement materials (including text, pictures and videos) consistent with brand images, and optimize the materials in real time according to the data in the delivery process. The invention creatively provides an ASA advertisement keyword and material generation system based on AI. The system utilizes strong natural language understanding capability and XLNet advantages in the aspects of text generation and multi-mode information processing, efficiently analyzes advertisement demands of advertisers, accurately grasps product selling points and audience characteristics, and intelligently constructs advertisement keywords and materials conforming to branding. On the basis, the invention designs a set of dynamic optimization algorithm, counts advertisement effect data in real time, and continuously optimizes the combination strategy of keywords and the presentation mode of creative. Experimental results show that the method can remarkably improve key effect indexes such as click rate and conversion rate of ASA advertisements, greatly improve automation degree and efficiency of advertisement optimization while improving throwing effect, and provide a new break for ASA advertisement practice.

Disclosure of Invention

In order to solve the above problems, the present invention provides an artificial intelligence based ASA advertisement keyword and material generating system, the structure diagram of which is shown in fig. 1, and the system comprises the following modules:

and the ASA advertisement demand understanding module is used for completing the understanding of advertisement demand according to the following steps:

Step 1.1, adopting a BiLSTM-CRF architecture deep learning model to analyze the requirements of an advertiser;

step 1.2 introduction of multi-head attention mechanism MHA enhanced advertising intent understanding:

Q,K,V=H^Q,K,VW^Q,K,V;

MHA(Q,K,V)=concat(head₁,...,head_k)W^O;

wherein H ^Q,K,V represents the input feature matrix of the advertisement demand, W ^Q,K,V represents the learnable weight matrix, d represents the dimension of the attention header, head _k represents the output of the kth attention header, concat represents the vector concatenation operation, W ^O represents the output mapping matrix, and Q, K, V represent the query, key, and value matrices, respectively.

And the ASA advertisement keyword generation module is used for generating advertisement keywords according to the following steps:

step 2.1, mining accurate advertisement keywords based on knowledge graphs:

Wherein, A knowledge sub-graph representing the advertising product p, (h, r, t) represents head, relationship and tail entity triples in the sub-graph, v _h,v_r,v_t represents their vector representations, respectively, |·|| ₂ represents the L2 norm, and argmin represents the parameter value that minimizes the objective function.

Step 2.2 advertisement keyword expansion based on XLNet:

P(w_i|S_\i)＝XLNet(S|w_i=[MASK]);

Where S represents the advertisement seed keyword sequence, w _i represents the ith word, h _i represents the hidden state of the last layer of position w _i of XLNet, Q (w _i) represents the advertisement keyword quality assessment score, b _i and b _j represent bias terms, w _i and w _j represent word vectors, S _\i represents a sequence other than the ith word, and [ MASK ] represents a MASK mark.

And the ASA advertisement material generation module is used for generating advertisement materials according to the following steps:

Step 3.1, designing a multi-mode transducer encoder to process advertisement materials:

h^(m)＝MMTE(x^(m),p^(m)),m∈{t,i,v};

Wherein x ^(m) represents a content feature vector of the advertisement material in a mode m, p ^(m) represents a position feature vector, m represents a mode type, t, i and v represent a text mode, an image mode and a video mode respectively, h ^(m) represents an output mode feature, and MMTE represents a multi-mode transducer encoder.

Step 3.2 introducing cross-modal matching loss optimization ad creative:

Wherein, Representing a collection of ad samples within a batch,Representing a characteristic representation of an advertisement sample x in a modality m, x ⁺ represents a positive sample matching sample x, x' represents other samples in batch, and represents a dot product operation.

The advertisement effect estimating module is used for estimating the advertisement effect according to the following steps:

step 4.1, modeling advertisement creatives and keywords by a design feature interaction layer:

z_ad＝MLP(e_ad)⊙v_ad;

z_kw＝MLP(e_kw)⊙v_kw;

z=[z_ad,z_kw,z_ad⊙z_kw];

Wherein e _ad and e _kw are Embedding vectors of the ad creative and the keywords, v _ad and v _kw are additional feature vectors of the ad creative and the keywords, respectively, and by-Hadamard product, [. Cndot.,. Cndot. ] represents vector concatenation and MLP represents a multi-layer perceptron.

The advertisement delivery optimization module is used for carrying out advertisement delivery optimization according to the following steps:

step 5.1, using a hierarchical multi-arm slot machine strategy (HMAB) to optimize advertisement delivery:

Where a represents the advertisement placement strategy, EU _PK (a) represents the expected utility calculated based on a priori knowledge, n _a represents the number of choices of strategy a, n _i represents the number of choices of the ith strategy, alpha represents the exploration factor (for balance exploration and utilization), Representing the total number of choices for all policies.

Step 5.2, optimizing advertisement bid by adopting evolution algorithm:

Fitness(b)＝ω₁×CTR_b+ω₂×Rank_b-ω₃×CPC_b+ω₄×Quality_b;

Where b represents the advertisement bid vector, CTR _b represents the estimated click rate, rank _b represents the advertisement ranking score, CPC _b represents the cost per click, quality _b represents the advertisement Quality score, and ω ₁,ω₂,ω₃,ω₄ represents the click rate, ranking, cost and weight coefficients of Quality, respectively.

The invention also provides a storage medium for generating computer-implemented instructions based on the AI-based ASA advertisement keywords and materials, the storage medium can be read by computer equipment, and the stored computer program instructions can realize the full-flow intelligent processing of ASA advertisements when being executed by a processor. The storage medium comprises a deep understanding module for the demand of an advertiser, the deep learning module is used for carrying out demand analysis, the multi-head attention mechanism MHA is used for enhancing advertisement intention understanding, an input feature matrix of the advertisement demand is constructed, and feature processing is carried out through a learnable weight matrix, so that accurate grasp and structural expression of the advertisement demand are realized.

The core function module of the storage medium comprises an advertisement keyword generation system based on a knowledge graph and XLNet, a material processing engine of a multi-mode transducer, an effect prediction framework of multi-task learning and a throwing optimization strategy of a layered multi-arm slot machine. The keyword generation system performs accurate keyword mining by constructing a product knowledge subgraph, dynamic expansion of keywords is achieved by utilizing XLNet models, a material processing engine can uniformly process multi-mode contents such as texts, images and videos, generation quality of advertisement creatives is optimized through cross-mode matching loss, an effect prediction framework adopts a feature interaction layer to model relevance between the advertisement creatives and the keywords, multiple effect indexes are predicted through multi-task learning, and a delivery optimization strategy comprehensively utilizes HMAB algorithm and evolution algorithm to achieve multi-level intelligent optimization and bid adjustment from Campaign to Ad Group.

To ensure efficient operation of the system, the storage medium also contains a complete data management and system monitoring module. The data management module is responsible for establishing an index structure of the advertisement material library, storing and maintaining entity relation data, model parameters and training data of the knowledge graph, and supporting incremental update of the data and a garbage recycling mechanism. And the system monitoring module records the running state and performance index of the system in real time, monitors the prediction effect of the model, and self-adaptively adjusts the system parameter configuration to realize abnormality detection and fault recovery. The storage medium may be various forms of computer readable storage medium such as ROM, RAM, magnetic disk, optical disk, etc., and is provided with a special area for storing data such as intermediate calculation results and evaluation indexes, and a buffer area for temporary data storage during program execution, so as to ensure stability and efficiency of system operation.

According to the method, through the innovative combination of the knowledge graph and XLNet, the accuracy and coverage range of keyword generation are remarkably improved, meanwhile, unified processing and generation of text, images and video materials are realized by adopting a multi-mode transducer architecture, and semantic consistency among different forms of materials is ensured through cross-mode matching loss. Based on the effect prediction framework of multi-task learning, the problem of correlation modeling among the prediction tasks is solved through a feature interaction layer and an uncertainty weighting mechanism, and high accuracy is still maintained in a data sparse scene.

On the system level, the hierarchical multi-arm slot machine (HMAB) optimization strategy realizes multi-level intelligent optimization from Campaign to Ad Group, and the automation degree and budget utilization efficiency of advertisement delivery are remarkably improved by combining bid optimization of an evolution algorithm. The whole system greatly reduces the manual intervention cost through end-to-end automatic processing, maintains stronger market adaptability, and can quickly adjust strategies according to put data.

Drawings

Fig. 1 is a system configuration diagram of the present application.

Detailed Description

XLNet is a tranformer language model proposed by google, and achieves better performance than BERT on multiple NLP tasks. Unlike traditional language models, XLNet adopts the training paradigm of PermutationLanguageModeling, which can model the bi-directional dependency of text sequences, and has stronger semantic understanding and context awareness. By applying XLNet to automatic generation of advertisement text, semantic information in a material library can be fully mined, language styles and industry expressions of brands are learned, and then smooth, natural and powerful advertisement expressions are generated.

Advertisement optimization needs to comprehensively consider material information of different modes such as texts, pictures and videos. XLNet adopts a Two-STREAMSELF-attribute mechanism, and can flexibly fuse the characteristic representations of different modes. The mechanism maps the inputs of different modalities to the same semantic space by introducing an identification vector (ModalityIndicator) and a location vector (PositionEmbedding), thereby modeling interactions and associations between modalities. On the basis, the invention designs a Multi-mode transducer encoder (Multi-modalTransformerEncoder, MMTE) which realizes the end-to-end representation learning of text, image and video characteristics:

h^(m)＝MMTE(x^(m),p^(m)),m∈{t,i,υ} (1);

Wherein x represents the input content feature vector, p represents the input position feature vector, and t, i and v represent text, image and video modes respectively. MMTE adaptively fusing information of different modes through a Self-attribute mechanism, and finally obtaining a unified multi-mode characteristic representation h. Based on this representation, the present invention further designs a Cross-modal matching (CM) loss function:

Wherein the method comprises the steps of Is a collection of advertisement samples within a batch. Equation (2) explicitly models cross-modal consistency of advertisement material by maximizing the degree of matching of different modal features within the same advertisement.

The intelligent ASA advertisement optimizing system mainly comprises a demand understanding module, a keyword generating module, a material generating module and an effect estimating module. The system comprises a demand understanding module, a keyword generation module, a material generation module, an effect prediction module and a machine learning algorithm, wherein the demand understanding module and the keyword generation module realize the structural representation of advertisement demands and the automatic generation of keywords, the material generation module comprehensively utilizes XLNet text generation and multi-mode feature fusion capability to automatically generate high-quality image-text and video materials, and the effect prediction module predicts key effect indexes of advertisement creative and keyword combination through the machine learning algorithm based on historical data. The four modules work cooperatively to finally output the optimized advertisement delivery configuration.

The demand understanding module is intended to translate advertising demand provided by advertisers in natural language into a structured information representation. Traditional demand understanding mainly relies on manually designed keyword extraction rules and template matching, and has limited generalization capability. The invention combines advanced natural language processing technology to design a set of end-to-end advertisement demand understanding model based on deep learning.

The backhaul network of the model adopts BiLSTM-CRF architecture. The BiLSTM layer models the context information of the input text sequence, and the CRF layer further considers the constraint relation among the slot labels to decode the optimal slot filling path. In addition, the present invention introduces a Multi-head attention (Multi-HeadAttention, MHA) layer before BiLSTM layers:

the MHA layer learns different semantic views of the input sequence in parallel through a plurality of attention heads, and accuracy of intention recognition and slot filling is improved.

The model is pre-trained on a large-scale advertisement corpus, and a set of ASA advertisement specific intention slot labeling system is designed according to a business scene. For the given advertisement demand description, the model can accurately analyze the key information such as product types, audience characteristics, popularization targets and the like, and output normalized JSON format representation. The structural requirement representation can effectively support links such as subsequent keyword expansion, material generation and the like.

High quality keywords are the basis for ASA advertising. The invention provides a keyword generation method integrating a knowledge graph and deep learning. The method comprises two main steps of targeted keyword mining and generalized keyword expansion.

Firstly, the invention constructs a product knowledge graph by utilizing a commercial large model based on product information provided by an advertiser. The knowledge graph uses the product as a central node, and related concepts are linked through various semantic relations (such as attributes, functions, usage scenes and the like). The invention adopts a map embedding method based on TransE to learn the low-dimensional vector representation of the product nodes:

Wherein the method comprises the steps of And (h, r, t) is a triplet in the subgraph, and represents a head entity, a relation and a tail entity respectively. TransE assume v _h+v_r≈v_t that learns the embedded representation of entities and relationships by minimizing translational errors. On the basis, the invention designs a heuristic scoring mechanism, comprehensively considers the factors such as semantic relevance of keywords and products, industry popularity, historical click rate and the like, and excavates a group of highly targeted seed keywords from the knowledge graph.

The seed keywords embody accurate grasp of the selling points of the products, but the quantity is limited, so that the coverage requirement of advertisement delivery is difficult to meet. Therefore, in the second step, more related words are generated through expanding generalized keywords on the basis of the seed keywords. Inspired by MaskLanguageModel, MLM paradigms, the method adopts XLNet as a generator, takes seed keywords as input, and generates generalized keywords in an iterative mask prediction mode. Specifically, the S= { w ₁,w₂,...,w_K } is taken as a seed keyword sequence, the alpha proportion words are masked randomly, and the probability distribution of the masked words is obtained by inputting XLNet:

P(w_i|S_\i)＝XLNet(S|w_i=[MASK]);

Based on the distribution, new words w '_i are sampled to replace the original mask words, a new sequence S' is obtained, and the process is repeated until a sufficient number of generalized keywords are generated. In order to ensure the quality of the generated words, the invention sets constraint conditions such as top-k sampling and the like, introduces a quality estimator at the last layer of XLNet, and dynamically adjusts the weight of the generated words:

h _i is the hidden state of the position w _i of the last layer XLNet, and Q (w _i) is the w _i mass fraction output by the quality estimator. The generalized keywords obtained by the method are highly related to seed keywords semantically, and meanwhile, the form is richer, so that the method is beneficial to attracting the extensive interests of potential users.

The method improves the coverage of the keywords and simultaneously well ensures the relevance by integrating the targeted keywords and the generalized keywords. The application of a large number of real advertisement cases shows that the keyword generation method based on the knowledge graph + XLNet can remarkably improve the display quantity and click rate of ASA advertisements.

High quality advertising creatives are key to improving advertising effectiveness. Traditional material designs rely primarily on experienced advertising creative personnel, with high costs and low efficiency. The invention provides a multi-mode and multi-task end-to-end material generation method which can automatically generate highly personalized graphics and video materials according to an advertisement corpus.

The method takes XLNet as the basis of a generation model and is integrated with a specific design facing ASA material generation task, 1) advertisement industry corpus is introduced in a pre-training stage, understanding and generation capacity of the advertisement text by the model is enhanced, 2) a multi-mode transducer encoder (formula (1) is taken as a bottom layer feature extractor to realize unified modeling of image-text and video features, 3) a plurality of decoders are designed to respectively complete different material generation subtasks (such as text, key frames and video clips) and collaborative training among the decoders is realized through cross-mode matching loss (formula (2), and 4) an external knowledge injection mechanism is added in the decoders, and the material generation is guided by using priori information such as a knowledge map.

During model training, the invention constructs a large-scale advertisement material corpus (comprising text, pictures and videos) into a series of pairs of keywords and advertisement originality, and the keyword information is transmitted to each decoder through a concentration mechanism to guide the generation of corresponding materials. For example, for a document generation task, the goal of the model is to maximize the conditional probability as follows:

Wherein y is the generated text sequence, x is the input keyword sequence, Is a knowledge graph corresponding to the key words. Similarly, the keyframe generation task generates a task with x and x sumsFor input, a key frame sequence is generated by autoregressive decoding. The video editing task is further based on the key frames, and the shot assembly is optimized through reinforcement learning, so that the video is optimal in information quantity, attractive appearance, rhythmic sense and the like.

According to the method, the multi-mode material generation model is trained on massive advertisement material corpus, and a set of evaluation system based on manual scoring is designed to evaluate the generation quality. The offline experiment and the online A/B test result show that the model can automatically design personalized advertising creatives which are highly attractive to users according to the input keywords, and the average click rate and the conversion rate are both remarkably improved. Meanwhile, the model has good generalization capability, can generate high-quality materials for new products and new scenes, and saves a great amount of creative design cost for advertisers.

The advertisement effect prediction is important to the optimization of advertisement delivery strategies. The invention provides a Multi-TASKLEARNING MTL-based depth effect prediction model, which can simultaneously predict key indexes such as click rate (CTR), conversion rate (CVR), retention rate (RetentionRate, RR) and the like of advertisement creative and keyword combination.

The model takes ad creatives (ad) and keywords (kw) as inputs and is first mapped to dense vectors through respective Embedding layers. Because of the diversity of creative and keyword types, the invention further introduces FeatureInteraction layers to model the high-order interaction features of the creative and keyword types:

zad=MLP(ead)⊙vad zkw=MLP(ekw)⊙vkw z=[zad,zkw,zad⊙zkw];

e _ad and e _kw are Embedding vectors of the creative and the keyword, respectively, and as Hadamard product. By introducing the MLP transform and the second order interaction term z _ad⊙z_kw, the layer can flexibly learn the nonlinear relationship between features.

The output z of FeatureInteraction layers is transmitted into a plurality of Task-specific MLP classifiers to respectively predict different target indexes:

These classifiers share underlying features z, so that correlations between different metrics can be explicitly modeled while maintaining certain task variability through independent classification layers. The model training process adopts a joint learning paradigm, and combines the loss of a plurality of tasks:

L=λ₁L_CTR+λ₂L_CVR+λ₃L_RR;

lambda _i is the weight coefficient of each task. In addition, the invention also introduces a Gate mechanism after each classifier to adaptively adjust task weights:

Wherein the method comprises the steps of Is a learnable parameter vector. The Gate mechanism enables the model to dynamically allocate task weights according to the input creative and keywords, and flexibility of prediction is improved.

The MTL-based effect prediction model fully utilizes the correlation among different indexes, and can accurately predict the advertisement effect in a sparse scene. The offline experiment shows that the AUC value of the independent prediction model is improved by 3% -5% compared with that of the traditional independent prediction model.

The goal of advertisement delivery optimization is to find keyword combinations and creative strategies to maximize the overall revenue of the advertisement in the event that budget constraints are met. The invention provides a hierarchical delivery optimization framework, which optimizes the combination strategy, bid and budget allocation of keywords and creatives from top to bottom.

The top layer uses a Multi-arm slot machine (Multi-armedBandit, MAB) model, treating each keyword or creative as a slot machine arm. As the delivery process proceeds, the MAB explores the expected benefits of the different arms through e-greedy and uses UCB algorithm to balance exploration and utilization, gradually converging to globally optimal delivery combinations.

Based on Multi-armedBandit and MAB, the invention designs a series of improvements by considering specific constraint and target of ASA scene, 1) introducing a hierarchical slot machine structure (HierarchicalMAB, HMAB), optimizing budget allocation at Campaign level, optimizing keyword and creative combination at Ad Group level, improving optimization efficiency, 2) designing a reward function based on expected utility (ExpectedUtility, EU), and considering multiple factors such as click rate, conversion rate, cost and the like, and reasonably weighing advertising benefits and cost, 3) adding priori Knowledge (Prior knowledges, PK) guide in bandit algorithm, fully utilizing information such as click rate provided by an effect estimation model, and accelerating strategy convergence. The revised HMAB rewards may be expressed as:

Where a is some combination policy, EU _PK (a) calculates the expected utility based on a priori knowledge, n _a is the number of times policy a has been selected, n _i is the number of selections of the ith arm, and α is the exploration factor.

After Multi-armedBandit, the MAB gets the combined strategy, the lower layer uses an evolution algorithm (EvolutionaryAlgorithm, EA) to further optimize the bidding of the keywords. Unlike traditional manual bidding rules, EA iteratively searches for optimal bids by selecting, crossing, mutating, etc., from a plurality of randomly generated bidding schemes. Wherein, the Fitness function comprehensively considers factors such as advertisement click rate, ranking, click unit price and the like:

b is a bid vector, CTR _b、Rank_b is a click rate and ranking estimated based on an effect estimation model respectively, CPC _b is a click unit price, quality _b is an advertisement material Quality score, and omega _i is a weight coefficient. After each iteration, higher scoring bidding schemes are preserved and cross-variation is performed, low quality schemes are eliminated, and the multiple rounds of evolution are performed until bidding converges.

In order to verify the effectiveness of the method provided by the invention, the method provided by the invention performs large-scale experimental comparison and analysis in a real ASA advertisement putting scene. The experimental data set covers the delivery data of a plurality of advertisers in a plurality of industries, and the time span is 3 months.

The invention adopts an A/B test framework to carry out experiments, and the flow of each advertiser is randomly divided into a control group (ControlGroup, CG) and an experimental group (TG), wherein the flow ratio of the control group to the experimental group is 1:1. The control group adopts the original delivery strategy of the advertiser, and the experimental group adopts the optimization method provided by the invention. The effect of the improvement of the experimental group compared with the control group was observed.

The evaluation indexes used in the experiment mainly comprise:

the baseline model for experimental comparison included:

1) Classical TF-IDF keyword extraction and Manual configuration (TF-IDF+Manual)

2) Rule template-based material generation (Template Creation)

3) DeepFFM click-through rate estimation model of independent task (DeepFFM)

The invention compares the effect of the knowledge graph + XLNet (KG + XLNet) keyword generation method and the TF-IDF+manual baseline, and the result is shown in Table 1:

Table 1:

it can be seen that the KG+ XLNet method achieves remarkable effect improvement in various industries, and average CTR and CTR are respectively and greatly improved. The invention considers that the main reasons are that on one hand, the product updating iteration of the two industries is fast, the homogenization competition is strong, the accurate touch requirements of advertisers are strong, on the other hand, the knowledge graph can better mine the subdivision attribute of the product, and the accurate and various keyword combinations are obtained by combining the strong generating capacity of XLNet. The improvement of the education industry is relatively small because users in the industry have more stable searching habits, and meanwhile, the subjective preference of advertisers on keywords is stronger, so that the optimization space of the algorithm is limited to a certain extent. In general, the KG+ XLNet method can fully integrate industry knowledge and language priori, automatically generate customized keywords and bring substantial effect improvement to advertisers.

The invention compares the effect of the multimodal XLNet (MM-XLNet) material generation model with the Template Creation base line. Since Template Creation relies on manual configuration, the workload is large, 3 representative industries are selected for comparison, and the results are shown in table 2:

Table 2:

Experimental results show that the MM-XLNet model is superior to the traditional template-based method in generation of pictures and video materials, and further analysis shows that the materials generated by the MM-XLNet are superior in visual attraction and information integrity, and the generated materials are more fit for the requirements of users due to the deep understanding of the model on images and texts. In addition, through fusion modeling of multi-mode information, the MM-XLNet can better grasp the internal connection of different material forms, so that the generated picture video is highly matched with the brand adjustment of advertisers and the selling point of products.

The invention also establishes a multi-task learning (MTL) effect estimation model provided by the invention, and compares the model with a traditional independent task DeepFFM model, and the result is shown in a table 3. Two learning paradigms, single task Loss (STL Loss) and multiple task joint Loss (MTL Loss), were used in the experiments, respectively. It can be seen that the model adopting the MTL learning paradigm achieves better effects no matter the CTR or CVR is estimated, and the joint learning is helpful for sharing and transferring knowledge among different tasks. In addition, the dynamic MTL Loss and the simpler static weighting MTL Loss added with the Gate mechanism are further improved, and the importance difference of the Gate mechanism in modeling different tasks and the value of improving model generalization are proved.

Table 3:

To further verify the superiority of the MTL model, the present invention has counted the overall effectiveness of advertisement delivery after using different effectiveness prediction models, as shown in Table 4. It can be seen that whether the HMAB optimized or manually configured release strategy is adopted, the effect prediction based on the MTL model (Ours) leads the overall ROI and CPA to be better than the DeepFFM model, and the effect improvement caused by modeling the connection among different indexes and relieving the data sparsity in the multi-task learning is again proved. It is worth mentioning that HMAB the optimization framework has obvious advantages over manual rules in terms of various indexes, which benefits from HMAB being able to continuously learn and dynamically adjust strategies, and fully utilizes the effect prediction result to guide the decision process of advertisement delivery.

Table 4:

In conclusion, the effectiveness of a series of optimization methods proposed by the invention is confirmed by large-scale on-line experimental results. The knowledge map is combined with XLNet to enhance the richness and accuracy of keyword expansion, the multi-mode XLNet material generation model provides personalized and diversified high-quality advertising creatives, the MTL effect prediction model remarkably improves the prediction effects of key indexes such as click rate and conversion rate, and the hierarchical MAB optimization framework realizes the self-adaptive adjustment of strategies to maximize advertising benefits. The comprehensive application of the methods realizes the intellectualization and automation of each link of advertisement delivery, and the input-output Ratio (ROI) of an advertiser is greatly improved.

Aiming at three core problems of keyword expansion, material generation and effect prediction in ASA advertisement optimization, the invention provides an end-to-end solution based on knowledge graph and multi-mode learning. The scheme creatively fuses the natural language understanding capability, the structural expression mode of the knowledge graph, the text generation of XLNet and the multi-modal feature modeling advantages, and achieves remarkable effect improvement in massive advertisement putting scenes.

Specifically, the main contributions of the present invention are as follows:

1) The method comprises the steps of introducing a knowledge graph into ASA advertisement optimization, carrying out explicit modeling on priori knowledge such as product attributes and user characteristics based on the knowledge graph, and learning vectorization representation of products through a graph neural network so as to realize automatic mining of targeted keywords.

2) By improving XLNet models, the end-to-end generation of multi-mode materials such as advertisement documents, pictures, videos and the like is realized through means of iterative mask prediction, multi-mode feature fusion, cross-mode consistency constraint and the like. The generated personalized creative effectively matches product selling points and user preferences, and improves the attractive force and conversion effect of advertisements.

3) The effect prediction framework under the multitask learning paradigm is provided, the advertisement creative and the high-order interaction of keywords are modeled through Feature Interaction layers, the weights of different task targets are dynamically adjusted by introducing a Gate mechanism, and the predicted effect of key indexes such as click rate, conversion rate and the like is greatly improved in a sparse data environment.

4) And a HMAB layering delivery optimization strategy is designed, and intelligent advertisement budget allocation and delivery combination optimization is realized at Campaign and Ad Group different granularities. The framework is highly self-adaptive, the strategy can be adjusted in real time according to the environmental change, the effect estimated information is fully utilized to guide the delivery decision, and the ROI of the advertiser is maximized.

Claims

1. The AI-based ASA advertising keyword and material generation system is characterized in that the system includes the following modules:

ASA advertising demand understanding module;

ASA advertising keyword generation module;

ASA creative generation module;

Advertising effect estimation module;

Advertisement delivery optimization module.

2. The AI-based ASA advertising keyword and material generation system according to claim 1 is characterized in that the ASA advertising demand understanding module understands the advertising demand according to the following steps:

Step 1.1 Use the deep learning model of BiLSTM-CRF architecture to analyze advertiser needs;

Step 1.2 introduces the multi-head attention mechanism MHA to enhance the understanding of advertising intent:

Q, K, V = H ^{Q, K, V} W ^{Q, K, V} ;

MHA(Q,K,V)=concat(head ₁ ,...,head _k )W ^O ;

Among them, H ^{Q, K, V} represents the input feature matrix of advertising demand, W ^{Q, K, V} represents the learnable weight matrix, d represents the dimension of the attention head, head _k represents the output of the kth attention head, concat represents the vector concatenation operation, W ^O represents the output mapping matrix, Q, K, V represent the query, key and value matrices respectively.

3. The AI-based ASA advertising keyword and material generation system according to claim 1, wherein the ASA advertising keyword generation module generates advertising keywords according to the following steps:

Step 2.1 Precise advertising keyword mining based on knowledge graph:

in, represents the knowledge subgraph of the advertising product p, (h, r, t) represents the head entity, relation and tail entity triples in the subgraph, v _h , v _r , v _t represent their vector representations respectively, ||·|| ₂ represents the L2 norm, and arg min represents the parameter value that minimizes the objective function;

Step 2.2: Ad keyword expansion based on XLNet:

P(w _i |S _\i )=XLNet(S|w _i =[MASK]);

Among them, S represents the advertising seed keyword sequence, _wi represents the i-th word, _hi represents the hidden state of the last layer of XLNet at position w _i , Q( _wi ) represents the advertising keyword quality evaluation score, _bi and _bj represent bias terms, _wi and _wj represent word vectors, S _\i represents the sequence except the i-th word, and [MASK] represents the mask tag.

4. The AI-based ASA advertising keyword and material generation system according to claim 1, wherein the ASA advertising material generation module generates advertising materials according to the following steps:

Step 3.1 Design a multimodal Transformer encoder to process advertising materials:

h ^(m) = MMTE (x ^(m ), p ^(m) ), m∈{t, i, v};

Where x ^(m) represents the content feature vector of the ad material in mode m, p ^(m) represents the position feature vector, m represents the modality type, t _, i, v represent text, image, and video modalities respectively, h(m) represents the output modality feature, and MMTE represents the multimodal Transformer encoder.

Step 3.2 introduces cross-modal matching loss to optimize ad creatives:

in, Represents a set of advertising samples in a batch. represents the feature representation of the advertising sample x under modality m, x ⁺ represents the positive sample matching the sample x, x′ represents other samples in the batch, and · represents the dot product operation.

5. The AI-based ASA advertising keyword and material generation system according to claim 1, characterized in that the advertising effect estimation module estimates the advertising effect according to the following steps:

Step 4.1 Design the feature interaction layer to model advertising creatives and keywords:

_zad = MLP( _ead ) _⊙vad ;

z _kw =MLP(e _kw )⊙v _kw ;

z＝[z _ad ,z _kw ,z _ad ⊙z _kw ]; where e _ad and e _kw are the embedding vectors of advertising creativity and keywords respectively, _vad and v _kw are the additional feature vectors of advertising creativity and keywords respectively, ⊙ is the Hadamard product (element-wise product), [·,·] represents vector concatenation, and MLP represents multi-layer perceptron.

6. The AI-based ASA advertising keyword and material generation system according to claim 1, characterized in that the advertising delivery optimization module optimizes advertising delivery according to the following steps:

Step 5.1 Use the Hierarchical Multi-Armed Bandit Strategy (HMAB) for ad delivery optimization:

Where a represents the advertising strategy, EU _PK (a) represents the expected utility calculated based on prior knowledge, _na represents the number of selections of strategy a, _ni represents the number of selections of the ith strategy, and α represents the exploration factor used to balance exploration and exploitation. represents the total number of selections for all strategies;

Step 5.2: Use evolutionary algorithms to optimize advertising bids:

Fitness(b)=ω _l ×CTR _b +ω ₂ ×Rank _b -ω ₃ ×CPC _b +ω ₄ ×Quality _b ;

Among them, b represents the advertising bid vector, CTR _b represents the estimated click-through rate, Rank _b represents the advertising ranking score, CPC _b represents the cost per click, Quality _b represents the advertising quality score, ω ₁ , ω ₂ , ω ₃ , and ω ₄ represent the weight coefficients of click-through rate, ranking, cost, and quality, respectively.